Omri's AI Agent Marketplace

Personal marketplace of plugins, skills, commands, and utilities for Claude Code, Codex, Grok, and Cursor.

Installation

Each CLI gets its own native manifest rather than relying on another vendor's compatibility layer:

Runtime	Marketplace manifest	Per-plugin manifest
Claude Code	.claude-plugin/marketplace.json	.claude-plugin/plugin.json
Codex	.agents/plugins/marketplace.json	.codex-plugin/plugin.json
Grok	.grok-plugin/marketplace.json	.grok-plugin/plugin.json
Cursor / cursor-agent	.cursor-plugin/marketplace.json	.cursor-plugin/plugin.json

The examples below install or discover the x plugin; replace x with another plugin name from the marketplace.

Claude Code

From the plugin UI:

/plugins → Marketplace → Add → omriariav/omri-marketplace

Or from a terminal:

claude plugin marketplace add omriariav/omri-marketplace
claude plugin install x@omri-marketplace

Codex

codex plugin marketplace add omriariav/omri-marketplace
codex plugin list --marketplace omri-marketplace --available --json
codex plugin add x@omri-marketplace

Grok

grok plugin marketplace add omriariav/omri-marketplace
grok plugin list --available --json

Then open /marketplace or /plugins in Grok to install the desired plugin.

Cursor / cursor-agent

Cursor manages installation through Customize in the IDE, the interactive /plugin browser, or the team marketplace dashboard. The CLI can register and list marketplaces, but it does not yet expose a non-interactive plugin install command.

cursor-agent plugin marketplace add https://github.com/omriariav/omri-marketplace
cursor-agent plugin marketplace list --format json

For a team marketplace, import https://github.com/omriariav/omri-marketplace under Dashboard → Plugins. For local development, clone this repository and load the desired plugin directly:

cursor-agent --plugin-dir "$PWD/plugins/x"

Runtime behavior

All marketplace entries have native manifests on all four runtimes. Component behavior differs only where the hosts expose different primitives:

Plugin type	Claude Code	Codex	Grok	Cursor
Skill-based plugins	Native skills	Native skills	Native skills	Native skills
copy commands	Native commands + skills	Native skills	Native commands + skills	Native commands + skills
coacher frame	Session-start hook + skill	Manual skill	Session-start hook + skill	Always-on rule + skill
natbag snapshot	Hook + invocation fallback	Invocation fallback	Hook + invocation fallback	Invocation fallback

setup-pulse and claude-reviewer can be invoked from every runtime, but their purpose remains configuring or reviewing Claude Code.

Maintaining manifests

Claude marketplace metadata is the source of truth. Adding a plugin does not require hand-maintaining four catalogs:

1. Create plugins/<name>/.claude-plugin/plugin.json with a semantic version and add the plugin's components.
2. Add one entry to .claude-plugin/marketplace.json. name and local source are required; description, author, and keywords fall back to the plugin manifest when omitted.
3. Ensure the plugin has a Codex-compatible component: a skill, .mcp.json, or .app.json. For command-only plugins, the generator creates Codex skill wrappers automatically. Hook/rule/agent-only plugins need a hand-authored fallback skill because Codex does not ingest those Claude/Grok/Cursor primitives.
4. Regenerate and verify:

python3 scripts/sync-native-manifests.py
python3 scripts/sync-native-manifests.py --check

Display names, short descriptions, categories, and starter prompts are inferred for new plugins. Optional curated overrides live near the top of the generator; they are not required for onboarding.

Marketplace plugins are intentionally vendored in this repository: every source must be the local path ./plugins/<name>. External or object-form Claude marketplace sources are not supported because Codex, Grok, and Cursor catalogs are generated from the vendored plugin files.

Pull requests run the same --check command, so a new or updated plugin cannot merge with stale native catalogs.

For a plugin with Claude/Grok hooks, the generator creates an empty Cursor hook override to prevent incompatible hook discovery. Replace it with a non-empty Cursor-native .cursor-plugin/hooks.json when needed; regeneration preserves that hand-authored translation.

CI and releases

Every pull request to main runs the required CI check. It validates tracked JSON, Python, and shell files; exercises future-plugin generation and release fixtures; verifies generated marketplace catalogs; confirms that every native manifest has the same plugin name and semantic version; and runs real Codex, Grok, and Cursor CLI plugin smoke tests against pinned versions. A daily scheduled matrix repeats those smoke tests against each vendor's latest stable CLI so upstream compatibility breaks are detected without making pull requests depend on a moving target. Files ending in .json must contain strict JSON; use .jsonc for comment-bearing configuration where the consuming tool supports it.

Plugin releases are created from the source manifest, so adding a plugin does not require editing the release workflow:

1. Update plugins/<name>/.claude-plugin/plugin.json with the next semantic version.
2. Run python3 scripts/sync-native-manifests.py and commit all generated native manifests in the same pull request.
3. Merge the pull request after CI passes.
4. Open Actions → Release plugin → Run workflow, select main, and enter <name>.

The workflow re-runs validation, rejects duplicate releases, creates the canonical <name>-v<version> tag from main, and publishes a GitHub release with generated notes. New marketplace entries are discovered automatically by both CI and the release helper.

Migrating from omri-cc-stuff? GitHub redirects the old repository URL, but Claude Code and Codex plugin IDs include the marketplace name and do not migrate automatically. Register omri-marketplace, reinstall the same plugins under their new plugin@omri-marketplace IDs, verify them, and only then remove the old marketplace. Plugin names are unchanged; versions continue independently as features are added.

Commands

/copy:* - Format & Copy to Clipboard

Copy data/tables/text to clipboard, formatted for your destination.

Command	Description
/copy:slack	Format for Slack (ASCII tables, *bold*)
/copy:gchat	Format for Google Chat (same as Slack)
/copy:gmail	Format for Gmail/email (rich text paste)
/copy:gdocs	Format for Google Docs (Markdown, paste via Edit → Paste from Markdown)
/copy:jira	Format for Jira Cloud (Markdown)
/copy:md	Markdown (standard syntax)
/copy:cleartext	Plain text (universal)
/copy:richformat	Rich Text Format (RTF) for Word, Pages

Usage:

/copy:slack                    → Auto-detect what to copy, format for Slack
/copy:slack the summary table  → Copy specific content for Slack
/copy:gmail                    → Auto-detect, format as rich text for email

Features:

• Auto-detects what to copy from conversation context
• Formats text, tables, lists, code, summaries
• Numbers auto-formatted (e.g., $2.45M not 2450000)
• QuickChart URLs for numeric data visualization

Example:

You: Compare the top programming languages

Claude: | Language   | Stars  | Growth |
        |------------|--------|--------|
        | Python     | 142K   | +18%   |
        | JavaScript | 128K   | +12%   |
        | Rust       | 89K    | +31%   |

You: /copy:slack

Claude: ✓ Copied to clipboard (format: Slack)

/gdoc-math - Markdown + LaTeX → Google Doc with editable equations

Convert a Markdown file containing LaTeX math into a native Google Doc whose equations are real, editable equation objects — not images, not literal $$ text. This is the gap /copy:gdocs can't fill: Google Docs' "Paste from Markdown" silently drops $...$ math, and there's no Docs API to insert equations. The pipeline routes through .docx instead:

markdown+LaTeX  →[pandoc]→  .docx (OMML)  →[gws upload + convert]→  Google Doc

/gdoc-math ~/notes/laplace.md                 → convert a file
/gdoc-math ~/notes/laplace.md --name "Notes"  → set the Doc title
/gdoc-math                                    → use the last math content in the conversation

Requirements (the skill verifies all up front):

• pandoc — brew install pandoc
• python3 — parses gws output (default on macOS)
• gws (Google Workspace CLI), authenticated for Drive — go install github.com/omriariav/workspace-cli/cmd/gws@latest then gws auth login

Features:

• Inline ($…$) and display ($$…$$) math, plus \(…\) / \[…\] delimiters
• Equations land as native, clickable, editable Google Docs equation objects (verified via OMML round-trip)
• Auto-trashes the intermediate .docx — only the Google Doc remains
• Optional default_folder_id in config.json to drop Docs in a specific Drive folder
• Ships a runnable sample at examples/math-demo.md

Use /copy:gdocs for plain prose/tables with no math (clipboard, instant). Use /gdoc-math when the content has formulas that must stay editable.

/x - Post and Read on X (Twitter)

Two skills under one plugin, sharing the same X API credentials.

⚠️ Migrating from the old tweet plugin? This release renames the plugin (tweet → x) and moves its directory (plugins/tweet/ → plugins/x/). Claude Code's plugin cache still points at the old path, so /plugin update will fail with Plugin source not found at .../plugins/tweet. Uninstall the old tweet plugin via /plugin, then install x fresh. Your Keychain credentials (service x-api) are untouched and will be picked up automatically. See RELEASES.md for the full migration path.

/x:tweet - Post

Post tweets directly from Claude Code. Drafts the text, shows a preview with character count, and asks for approval before posting. Learns your voice over time from editing patterns and feedback.

/x:tweet Testing my new tweet skill from Claude Code
/x:tweet Announce the launch of our new attribution feature

Workflow:

1. Draft tweet text (or use your text as-is)
2. Preview with character count (max 280)
3. Approve / Edit / Cancel
4. Posts via X API v2, returns tweet URL

/x:read - Read

Fetch a tweet by URL or ID via the X API v2 and load it into the conversation as markdown — with author, timestamp, metrics, replied-to / quoted context, and media. Useful for "summarize this thread", "draft a reply to this", or quoting a tweet in another piece of writing.

/x:read https://x.com/jack/status/20
/x:read 20

Hand the resulting tweet ID to /x:tweet --reply-to <id> to chain a reply. Note: each read is metered by X (~$0.005 / post read on the standard tier).

Setup (one-time):

pip3 install requests-oauthlib

On first run, /x:tweet will pop up native macOS dialogs to store your X API credentials securely in Keychain (/x:read reuses the same credentials). Get keys from developer.x.com → your app → Keys and Tokens (Consumer Key/Secret + Access Token/Secret with Read+Write permissions).

Alternatively, set env vars in ~/.zshrc: X_API_KEY, X_API_SECRET, X_ACCESS_TOKEN, X_ACCESS_TOKEN_SECRET.

/skill-reviewer - Skill Design Evaluation

Evaluate any skill's design quality against best practices. Scores 10 dimensions (29-point rubric), detects anti-patterns, and optionally applies safe fixes.

/skill-reviewer publisher-lookup          → Review by skill name
/skill-reviewer --fix my-skill            → Review + auto-apply safe fixes
/skill-reviewer --compare skill-a skill-b → Side-by-side comparison
/skill-reviewer --with-logs my-skill      → Review + include transcript usage evidence

Dimensions scored: Progressive Disclosure, Description Quality, Gotchas, Non-Obvious Content, Flexibility, Setup & Config, Memory, Scripts, Frontmatter, Hooks Integration.

Grades: A (25-29), B (20-24), C (15-19), D (10-14), F (0-9)

Skill cleanliness signals (v1.1.0): Every per-skill review now includes a "Skill Cleanliness Signals" section adapting the skill-cleaner methodology by @steipete — description budget cost, duplicates across roots (with keep-priority), and opt-in usage evidence via --with-logs. Read-only; informs the rubric, doesn't replace it.

/find-session - Search Past Conversations

Search past Claude Code and Codex session history by keyword and get the right resume command (claude --resume <id> or codex resume <id>) to pick up where you left off. Results from both sources interleave by recency, each tagged with its source.

/find-session deploy                → Search current project (both sources) for "deploy"
/find-session --all taboola         → Search across all projects
/find-session --all --codex taboola → Codex sessions only
/find-session --all --json auth     → Structured JSON output

Features:

• Two sources in one search: Claude Code (~/.claude/projects/) and Codex (~/.codex/sessions/)
• --source claude|codex|both (default both), with shorthand --claude / --codex / --both
• Searches user messages and custom titles (not assistant text); Codex previews skip injected role: user boilerplate (# AGENTS.md, <environment_context>, etc.)
• Word-boundary matching for clean terms, literal match for punctuation (claude-mem, c++)
• Shows custom title (from /rename) when set, falls back to first message preview (Codex has no titles → always preview-based)
• Codex subagent sessions (guardian judges, spawned children) excluded by default; --include-subagents to keep them
• Optional Haiku summarization for untitled sessions (asks before running)
• --json mode for programmatic use (adds source + per-result resume_cmd)
• Git-aware project root resolution (works from subdirectories); Codex sessions scoped by recorded cwd

Configuration (config.json):

Setting	Default	Description
max_results	20	Max sessions to show
use_haiku_summary	true	Enable Haiku summarization prompt
haiku_threshold	50	Max results for Haiku to trigger
haiku_model	claude-haiku-4-5-20251001	Model for summarization
exclude_slugs	[]	Project slugs to skip in --all mode

/claude-reviewer - Claude Code Configuration Auditor

Review any project's .claude/ folder setup against best practices. Scores 8 dimensions (24-point rubric), detects anti-patterns, and produces an actionable improvement report.

/claude-reviewer                       → Review current project
/claude-reviewer ~/Code/myproject      → Review specific project
/claude-reviewer --global              → Review only ~/.claude/ global config
/claude-reviewer --with-logs           → Add transcript-based unused-skill detection

Skill fleet audit (v1.2.0): Every report now includes a ## Skill Fleet Audit section adapting the skill-cleaner methodology by @steipete at fleet level — total budget, trim candidates, duplicates across roots, and opt-in unused detection. Companion to skill-reviewer v1.1.0's per-skill signals.

Dimensions scored: CLAUDE.md Quality, Permission Hygiene, Modular Instructions, Custom Commands, Skills Setup, Agent Configuration, Git Hygiene, Progressive Disclosure.

Grades: A (21-24), B (16-20), C (11-15), D (6-10), F (0-5)

/natbag - Ben Gurion Airport (TLV) Flight Data

Live flight data, destination weather, and historical delay analysis for Ben Gurion Airport (TLV), pulling from data.gov.il (flights) and Open-Meteo (weather). A SessionStart hook quietly snapshots flight data so the historical database accumulates over time.

/natbag departures            → Departures board (status, gate, terminal)
/natbag arrivals              → Arrivals board
/natbag LY001                 → Status of a specific flight
/natbag delayed               → Only delayed / canceled flights
/natbag weather Paris         → Current weather at a destination city
/natbag history El Al         → On-time / delay history for an airline

Features:

• Bilingual (Hebrew + English) input and output
• Rolling ~3-day live window; gates, terminals, check-in zones
• Local SQLite history accrues daily snapshots → on-time performance, delay/cancellation rates, per-flight change tracking
• Ships IATA reference data (999 airlines, 9,240 airports); user history lives in ~/.natbag/
• Pure Python 3 stdlib + sqlite3 + curl — no API keys, no pip installs

coacher - Peer-Collaborator Frame at Session Start

Primes Claude with a peer-collaborator stance at SessionStart (based on Amanda Askell's prompting philosophy — positive framing, permission to push back, no apology spirals). A SessionStart hook injects the frame from frame.md; edit that file to tune the stance.

Command	Description
/coacher:status	Integrity check — verify the frame is in the current context
/coacher:audit	Cross-check frame.md against your CLAUDE.md and flag stance conflicts
/coacher:reset	Re-anchor the frame mid-session if Claude drifts into hedging
/coacher:rant <text>	Vent raw frustration → Claude extracts the intent and acts on it

Requires python3 (stdlib only). The hook degrades gracefully — if python3 is missing or the frame is empty, it exits cleanly with no failure.

Other Commands

Command	Description
/setup-pulse	Install claude-pulse statusline

/setup-pulse - Install Token Usage Statusline

Installs claude-pulse - shows real-time context usage in your statusline.

/setup-pulse

After running, you'll see: 72k/200k (36%) with color-coded warnings (green/yellow/red).

Contributing

Contributions welcome! Feel free to open a PR or issue.

License

Free to use with attribution. Credit: Omri Ariav